You are on page 1of 2
View Reviews Paper ID 1 Paper Title ‘A Network Embeddings based Recommendation Model with multi-factors consideration Track Name Paper Reviewer #4 Questions 1. Overall recommendation Accept 2. Detailed Comments Paper outlines using network embeddings based on multi factor matrix factorization to make better recommendations. Strengths: 1. Paper reads really well end to end and is easy to follow. Authors build up the sections in appropriate pace and technical detail 2. Empirical experimentation and results are great to see 3. The key contribution of combining network embeddings with user-item interaction matrix is interesting Weaknesses: 1. Are there other potential applications than for recommendation systems domain? Is there an opportunity to apply this approach to other types of problems and do we expect similar improvements? 2. This is not too much of an issue but cross domain learning seems to reduce down to adding more parameters and not very interesting to claim as a key contribution. However, the approach is interesting 3. Are there other benchmarks than Tmall that could be evaluated? Is Beibei data set used in one of the references ‘a good fit? 4. Itis unclear if authors have evaluated comparison methods themselves sicne the other methods dont all use accuracy, precision to evaluate model performance. It's unclear why the choice of metrics is appropriate as well Reviewer #2 Questions 1. Overall recommendation Strong Accept 2. Detailed Comments ~ This paper's main contribution is utlizing both implicit and explicit information to improve recommendations. - The authors propose mult-perceptron network embedding model to learn the representation of user and item. - Then the authors propose multi-factor matrix factorization to combine the network embeddings and user- interactions to predict user's preference for an item: - The experiments done on real-world datasets significantly outperforms the state-of-the-art recommendation algorithms demonstrating the effectiveness of the proposed approaches. ‘Comments: - The paper is well-written, and the contributions made in this paper are significant especially utilizing new information and proposing non-trivial algorithms such as multifactor matrix factorization. - The approaches proposed clearly advances the state-of-the-art in the space of recommendation systems. Reviewer #3 Questions 41. Overall recommendation Not Ready 2. Detailed Comments This paper proposes a framework to provide recommendations based on user web behaviors. The question studied in the paper is of first order importance to the industry and the proposed framework is innovative. However, there are many open questions in the design that has not been fully explored and thus | find the paper not ready for publication 1) A simple node2vec transformation on the network resulting information loss and is not fully representative of the user behavior. For example, the sequencing of the browsing order is lost in the structure. 2) The connection between item metadata and purchasing/browsing behaviors are not fully explored in the model. For example, meta data such as price, review, and demand are changing overtime. The current model setup does. not explore the joint variations between these two sets of data. 3) the author should spend more paragraphs describe the implementation details such as tuning selection and performance of model at each stage.

You might also like